Best Alternatives to LFM3 by Liquid AI in 2025
LFM3 by Liquid AI brings open-weight, 2B-class foundation models with a shared backbone across text, vision, audio, omni, and decision tasks, optimized for phones, laptops, and edge devices. Its 1.7B text model runs in about 1 GB at 4-bit precision with a 256K context, making it a strong pick for on-device AI. But depending on your hardware, licensing needs, ecosystem, or task focus, other small-model families may fit better. Here are the best alternatives to LFM3.
Llama 3.2 (1B and 3B)
Meta's compact Llama 3.2 models are the most widely adopted small open-weight option, with strong tooling across llama.cpp, Ollama, and mobile runtimes. They offer a mature ecosystem and broad community support, though they lack LFM3's unified multimodal backbone and 256K context.
Gemma 3 (1B and 4B)
Google's Gemma 3 small models deliver strong text quality with multimodal variants and a permissive license for many commercial uses. They are well supported on edge hardware and benefit from Google's optimization work, making them a solid alternative when you want proven quality and clear licensing.
Phi-4-mini and Phi-3.5-mini
Microsoft's Phi small models punch above their weight on reasoning, math, and coding for their size. They are ideal if your edge workload is reasoning-heavy rather than multimodal, and they run well in 4-bit on laptops and phones, though they trail LFM3 on audio and decision-model coverage.
Qwen2.5 (0.5B to 3B)
Alibaba's Qwen2.5 small models offer excellent multilingual performance, long context options, and a wide size range from 0.5B to 3B. They are a strong choice for global or multilingual apps and have broad quantization support, but they are primarily text-focused compared with LFM3's omni and audio models.
Mistral Small 3 and Ministral 3B/8B
Mistral's compact models are fast, Apache-2.0 licensed, and tuned for low-latency inference on edge and local hardware. They are a good fit when you need a permissive license and strong instruction-following, though they do not match LFM3's breadth across vision, audio, and decision tasks.
SmolLM2 (135M to 1.7B)
Hugging Face's SmolLM2 family targets ultra-small, on-device deployments with sizes down to 135M and strong benchmarks for its class. It is ideal for phones and microcontrollers where memory is the hard constraint, but it offers less multimodal capability and smaller context than LFM3.
MiniCPM 3 and MiniCPM-V
OpenBMB's MiniCPM models are designed for efficient edge deployment, with strong small-model text performance and capable vision-language variants for phones. They are a good alternative for on-device multimodal apps, though the ecosystem and licensing terms are less established than Llama or Gemma.
LFM3 stands out for its shared multimodal backbone and 256K context in a roughly 1 GB footprint, which few rivals match. If you prioritize ecosystem maturity, choose Llama 3.2 or Gemma 3; for reasoning-heavy edge tasks, Phi-4-mini; for multilingual apps, Qwen2.5; for permissive licensing and speed, Mistral Small 3 or Ministral; for the smallest possible footprint, SmolLM2; and for efficient on-device vision, MiniCPM. Evaluate each against your hardware, license, and modality requirements before committing.